bcsr_interface.c 13 KB

123456789101112131415161718192021222324252627282930313233343536373839404142434445464748495051525354555657585960616263646566676869707172737475767778798081828384858687888990919293949596979899100101102103104105106107108109110111112113114115116117118119120121122123124125126127128129130131132133134135136137138139140141142143144145146147148149150151152153154155156157158159160161162163164165166167168169170171172173174175176177178179180181182183184185186187188189190191192193194195196197198199200201202203204205206207208209210211212213214215216217218219220221222223224225226227228229230231232233234235236237238239240241242243244245246247248249250251252253254255256257258259260261262263264265266267268269270271272273274275276277278279280281282283284285286287288289290291292293294295296297298299300301302303304305306307308309310311312313314315316317318319320321322323324325326327328329330331332333334335336337338339340341342343344345346347348349350351352353354355356357358359360361362363364365366367368369370371372373374375376377378379380381382383384385386387388389390391392393394395396397398
  1. /* StarPU --- Runtime system for heterogeneous multicore architectures.
  2. *
  3. * Copyright (C) 2009-2016 Université de Bordeaux
  4. * Copyright (C) 2010, 2011, 2012, 2013, 2014, 2017 CNRS
  5. *
  6. * StarPU is free software; you can redistribute it and/or modify
  7. * it under the terms of the GNU Lesser General Public License as published by
  8. * the Free Software Foundation; either version 2.1 of the License, or (at
  9. * your option) any later version.
  10. *
  11. * StarPU is distributed in the hope that it will be useful, but
  12. * WITHOUT ANY WARRANTY; without even the implied warranty of
  13. * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
  14. *
  15. * See the GNU Lesser General Public License in COPYING.LGPL for more details.
  16. */
  17. #include <starpu.h>
  18. #include <common/config.h>
  19. #include <datawizard/coherency.h>
  20. #include <datawizard/copy_driver.h>
  21. #include <datawizard/filters.h>
  22. #include <datawizard/memory_nodes.h>
  23. #include <starpu_hash.h>
  24. #include <starpu_cuda.h>
  25. #include <starpu_opencl.h>
  26. #include <drivers/opencl/driver_opencl.h>
  27. /*
  28. * BCSR : blocked CSR, we use blocks of size (r x c)
  29. */
  30. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data);
  31. static const struct starpu_data_copy_methods bcsr_copy_data_methods_s =
  32. {
  33. .any_to_any = copy_any_to_any,
  34. };
  35. static void register_bcsr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface);
  36. static void *bcsr_handle_to_pointer(starpu_data_handle_t data_handle, unsigned node);
  37. static starpu_ssize_t allocate_bcsr_buffer_on_node(void *data_interface, unsigned dst_node);
  38. static void free_bcsr_buffer_on_node(void *data_interface, unsigned node);
  39. static size_t bcsr_interface_get_size(starpu_data_handle_t handle);
  40. static int bcsr_compare(void *data_interface_a, void *data_interface_b);
  41. static uint32_t footprint_bcsr_interface_crc32(starpu_data_handle_t handle);
  42. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size);
  43. struct starpu_data_interface_ops starpu_interface_bcsr_ops =
  44. {
  45. .register_data_handle = register_bcsr_handle,
  46. .allocate_data_on_node = allocate_bcsr_buffer_on_node,
  47. .free_data_on_node = free_bcsr_buffer_on_node,
  48. .copy_methods = &bcsr_copy_data_methods_s,
  49. .get_size = bcsr_interface_get_size,
  50. .interfaceid = STARPU_BCSR_INTERFACE_ID,
  51. .interface_size = sizeof(struct starpu_bcsr_interface),
  52. .footprint = footprint_bcsr_interface_crc32,
  53. .compare = bcsr_compare,
  54. .describe = describe,
  55. .handle_to_pointer = bcsr_handle_to_pointer
  56. };
  57. static void *bcsr_handle_to_pointer(starpu_data_handle_t data_handle, unsigned node)
  58. {
  59. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(data_handle, node));
  60. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *)
  61. starpu_data_get_interface_on_node(data_handle, node);
  62. return (void*) bcsr_interface->nzval;
  63. }
  64. static void register_bcsr_handle(starpu_data_handle_t handle, unsigned home_node, void *data_interface)
  65. {
  66. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface;
  67. unsigned node;
  68. for (node = 0; node < STARPU_MAXNODES; node++)
  69. {
  70. struct starpu_bcsr_interface *local_interface = (struct starpu_bcsr_interface *)
  71. starpu_data_get_interface_on_node(handle, node);
  72. if (node == home_node)
  73. {
  74. local_interface->nzval = bcsr_interface->nzval;
  75. local_interface->colind = bcsr_interface->colind;
  76. local_interface->rowptr = bcsr_interface->rowptr;
  77. }
  78. else
  79. {
  80. local_interface->nzval = 0;
  81. local_interface->colind = NULL;
  82. local_interface->rowptr = NULL;
  83. }
  84. local_interface->id = bcsr_interface->id;
  85. local_interface->nnz = bcsr_interface->nnz;
  86. local_interface->nrow = bcsr_interface->nrow;
  87. local_interface->firstentry = bcsr_interface->firstentry;
  88. local_interface->r = bcsr_interface->r;
  89. local_interface->c = bcsr_interface->c;
  90. local_interface->elemsize = bcsr_interface->elemsize;
  91. }
  92. }
  93. void starpu_bcsr_data_register(starpu_data_handle_t *handleptr, int home_node,
  94. uint32_t nnz, uint32_t nrow, uintptr_t nzval, uint32_t *colind,
  95. uint32_t *rowptr, uint32_t firstentry,
  96. uint32_t r, uint32_t c, size_t elemsize)
  97. {
  98. struct starpu_bcsr_interface bcsr_interface =
  99. {
  100. .id = STARPU_BCSR_INTERFACE_ID,
  101. .nzval = nzval,
  102. .colind = colind,
  103. .rowptr = rowptr,
  104. .nnz = nnz,
  105. .nrow = nrow,
  106. .firstentry = firstentry,
  107. .r = r,
  108. .c = c,
  109. .elemsize = elemsize
  110. };
  111. #ifndef STARPU_SIMGRID
  112. if (home_node == STARPU_MAIN_RAM)
  113. {
  114. STARPU_ASSERT_ACCESSIBLE(nzval);
  115. STARPU_ASSERT_ACCESSIBLE(nzval + nnz*elemsize*r*c - 1);
  116. STARPU_ASSERT_ACCESSIBLE(colind);
  117. STARPU_ASSERT_ACCESSIBLE((uintptr_t) colind + nnz*sizeof(uint32_t) - 1);
  118. STARPU_ASSERT_ACCESSIBLE(rowptr);
  119. STARPU_ASSERT_ACCESSIBLE((uintptr_t) rowptr + (nrow+1)*sizeof(uint32_t) - 1);
  120. }
  121. #endif
  122. starpu_data_register(handleptr, home_node, &bcsr_interface, &starpu_interface_bcsr_ops);
  123. }
  124. static uint32_t footprint_bcsr_interface_crc32(starpu_data_handle_t handle)
  125. {
  126. uint32_t hash;
  127. hash = starpu_hash_crc32c_be(starpu_bcsr_get_nnz(handle), 0);
  128. hash = starpu_hash_crc32c_be(starpu_bcsr_get_c(handle), hash);
  129. hash = starpu_hash_crc32c_be(starpu_bcsr_get_r(handle), hash);
  130. return hash;
  131. }
  132. static int bcsr_compare(void *data_interface_a, void *data_interface_b)
  133. {
  134. struct starpu_bcsr_interface *bcsr_a = (struct starpu_bcsr_interface *) data_interface_a;
  135. struct starpu_bcsr_interface *bcsr_b = (struct starpu_bcsr_interface *) data_interface_b;
  136. /* Two matricess are considered compatible if they have the same size */
  137. return ((bcsr_a->nnz == bcsr_b->nnz)
  138. && (bcsr_a->nrow == bcsr_b->nrow)
  139. && (bcsr_a->r == bcsr_b->r)
  140. && (bcsr_a->c == bcsr_b->c)
  141. && (bcsr_a->elemsize == bcsr_b->elemsize));
  142. }
  143. /* offer an access to the data parameters */
  144. uint32_t starpu_bcsr_get_nnz(starpu_data_handle_t handle)
  145. {
  146. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  147. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  148. #ifdef STARPU_DEBUG
  149. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  150. #endif
  151. return data_interface->nnz;
  152. }
  153. uint32_t starpu_bcsr_get_nrow(starpu_data_handle_t handle)
  154. {
  155. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  156. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  157. #ifdef STARPU_DEBUG
  158. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  159. #endif
  160. return data_interface->nrow;
  161. }
  162. uint32_t starpu_bcsr_get_firstentry(starpu_data_handle_t handle)
  163. {
  164. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  165. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  166. #ifdef STARPU_DEBUG
  167. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  168. #endif
  169. return data_interface->firstentry;
  170. }
  171. uint32_t starpu_bcsr_get_r(starpu_data_handle_t handle)
  172. {
  173. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  174. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  175. #ifdef STARPU_DEBUG
  176. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  177. #endif
  178. return data_interface->r;
  179. }
  180. uint32_t starpu_bcsr_get_c(starpu_data_handle_t handle)
  181. {
  182. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  183. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  184. #ifdef STARPU_DEBUG
  185. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  186. #endif
  187. return data_interface->c;
  188. }
  189. size_t starpu_bcsr_get_elemsize(starpu_data_handle_t handle)
  190. {
  191. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  192. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  193. #ifdef STARPU_DEBUG
  194. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  195. #endif
  196. return data_interface->elemsize;
  197. }
  198. uintptr_t starpu_bcsr_get_local_nzval(starpu_data_handle_t handle)
  199. {
  200. unsigned node;
  201. node = _starpu_memory_node_get_local_key();
  202. STARPU_ASSERT(starpu_data_test_if_allocated_on_node(handle, node));
  203. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  204. starpu_data_get_interface_on_node(handle, node);
  205. #ifdef STARPU_DEBUG
  206. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  207. #endif
  208. return data_interface->nzval;
  209. }
  210. uint32_t *starpu_bcsr_get_local_colind(starpu_data_handle_t handle)
  211. {
  212. /* XXX 0 */
  213. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  214. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  215. #ifdef STARPU_DEBUG
  216. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  217. #endif
  218. return data_interface->colind;
  219. }
  220. uint32_t *starpu_bcsr_get_local_rowptr(starpu_data_handle_t handle)
  221. {
  222. /* XXX 0 */
  223. struct starpu_bcsr_interface *data_interface = (struct starpu_bcsr_interface *)
  224. starpu_data_get_interface_on_node(handle, STARPU_MAIN_RAM);
  225. #ifdef STARPU_DEBUG
  226. STARPU_ASSERT_MSG(data_interface->id == STARPU_BCSR_INTERFACE_ID, "Error. The given data is not a bcsr.");
  227. #endif
  228. return data_interface->rowptr;
  229. }
  230. static size_t bcsr_interface_get_size(starpu_data_handle_t handle)
  231. {
  232. size_t size;
  233. uint32_t nnz = starpu_bcsr_get_nnz(handle);
  234. uint32_t nrow = starpu_bcsr_get_nrow(handle);
  235. uint32_t r = starpu_bcsr_get_r(handle);
  236. uint32_t c = starpu_bcsr_get_c(handle);
  237. size_t elemsize = starpu_bcsr_get_elemsize(handle);
  238. size = nnz*r*c*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  239. return size;
  240. }
  241. /* memory allocation/deallocation primitives for the BLAS interface */
  242. /* returns the size of the allocated area */
  243. static starpu_ssize_t allocate_bcsr_buffer_on_node(void *data_interface_, unsigned dst_node)
  244. {
  245. uintptr_t addr_nzval, addr_colind, addr_rowptr;
  246. starpu_ssize_t allocated_memory;
  247. /* we need the 3 arrays to be allocated */
  248. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface_;
  249. uint32_t nnz = bcsr_interface->nnz;
  250. uint32_t nrow = bcsr_interface->nrow;
  251. size_t elemsize = bcsr_interface->elemsize;
  252. uint32_t r = bcsr_interface->r;
  253. uint32_t c = bcsr_interface->c;
  254. addr_nzval = starpu_malloc_on_node(dst_node, nnz*r*c*elemsize);
  255. if (!addr_nzval)
  256. goto fail_nzval;
  257. addr_colind = starpu_malloc_on_node(dst_node, nnz*sizeof(uint32_t));
  258. if (!addr_colind)
  259. goto fail_colind;
  260. addr_rowptr = starpu_malloc_on_node(dst_node, (nrow+1)*sizeof(uint32_t));
  261. if (!addr_rowptr)
  262. goto fail_rowptr;
  263. /* allocation succeeded */
  264. allocated_memory =
  265. nnz*r*c*elemsize + nnz*sizeof(uint32_t) + (nrow+1)*sizeof(uint32_t);
  266. /* update the data properly in consequence */
  267. bcsr_interface->nzval = addr_nzval;
  268. bcsr_interface->colind = (uint32_t*) addr_colind;
  269. bcsr_interface->rowptr = (uint32_t*) addr_rowptr;
  270. return allocated_memory;
  271. fail_rowptr:
  272. starpu_free_on_node(dst_node, addr_colind, nnz*sizeof(uint32_t));
  273. fail_colind:
  274. starpu_free_on_node(dst_node, addr_nzval, nnz*r*c*elemsize);
  275. fail_nzval:
  276. /* allocation failed */
  277. return -ENOMEM;
  278. }
  279. static void free_bcsr_buffer_on_node(void *data_interface, unsigned node)
  280. {
  281. struct starpu_bcsr_interface *bcsr_interface = (struct starpu_bcsr_interface *) data_interface;
  282. uint32_t nnz = bcsr_interface->nnz;
  283. uint32_t nrow = bcsr_interface->nrow;
  284. size_t elemsize = bcsr_interface->elemsize;
  285. uint32_t r = bcsr_interface->r;
  286. uint32_t c = bcsr_interface->c;
  287. starpu_free_on_node(node, bcsr_interface->nzval, nnz*r*c*elemsize);
  288. starpu_free_on_node(node, (uintptr_t) bcsr_interface->colind, nnz*sizeof(uint32_t));
  289. starpu_free_on_node(node, (uintptr_t) bcsr_interface->rowptr, (nrow+1)*sizeof(uint32_t));
  290. }
  291. static int copy_any_to_any(void *src_interface, unsigned src_node, void *dst_interface, unsigned dst_node, void *async_data)
  292. {
  293. struct starpu_bcsr_interface *src_bcsr = (struct starpu_bcsr_interface *) src_interface;
  294. struct starpu_bcsr_interface *dst_bcsr = (struct starpu_bcsr_interface *) dst_interface;
  295. uint32_t nnz = src_bcsr->nnz;
  296. uint32_t nrow = src_bcsr->nrow;
  297. size_t elemsize = src_bcsr->elemsize;
  298. uint32_t r = src_bcsr->r;
  299. uint32_t c = src_bcsr->c;
  300. int ret = 0;
  301. if (starpu_interface_copy(src_bcsr->nzval, 0, src_node, dst_bcsr->nzval, 0, dst_node, nnz*elemsize*r*c, async_data))
  302. ret = -EAGAIN;
  303. if (starpu_interface_copy((uintptr_t)src_bcsr->colind, 0, src_node, (uintptr_t)dst_bcsr->colind, 0, dst_node, nnz*sizeof(uint32_t), async_data))
  304. ret = -EAGAIN;
  305. if (starpu_interface_copy((uintptr_t)src_bcsr->rowptr, 0, src_node, (uintptr_t)dst_bcsr->rowptr, 0, dst_node, (nrow+1)*sizeof(uint32_t), async_data))
  306. ret = -EAGAIN;
  307. _STARPU_TRACE_DATA_COPY(src_node, dst_node, nnz*elemsize*r*c + (nnz+nrow+1)*sizeof(uint32_t));
  308. return ret;
  309. }
  310. static starpu_ssize_t describe(void *data_interface, char *buf, size_t size)
  311. {
  312. struct starpu_bcsr_interface *bcsr = (struct starpu_bcsr_interface *) data_interface;
  313. return snprintf(buf, size, "b%ux%ux%ux%ux%u",
  314. (unsigned) bcsr->nnz,
  315. (unsigned) bcsr->nrow,
  316. (unsigned) bcsr->r,
  317. (unsigned) bcsr->c,
  318. (unsigned) bcsr->elemsize);
  319. }